Have a "revert to factory default" setting for that resets LBT to point to the BLAS that shipped with Julia (for 1.6 and earlier)

[carstenbauer]

Is it possible to switch back to OpenBLAS? If this is a feature it should be documented.

I see that there is a enable_openblas_startup() function in install.jl: https://github.com/JuliaComputing/MKL.jl/blob/b6283ca3def34ee6329023e8e8ec2700c2328ff8/src/install.jl#L104

How should one call this function to revert the changes?

[carstenbauer]

I (seem to have) managed to revert back to openblas by setting ENV["USEBLAS64"]=true (otherwise there is an error) and changing the line enable_mkl_startup() to enable_openblas_startup(). Afterwards I have

julia> BLAS.vendor()
:openblas64

again.

Would be nice to have this work without having to adjust build.jl.

[ViralBShah]

I revert by deleting and reinstalling Julia - since it doesn't affect the packages or anything. I'd rather get the default build from julialang.org instead of building my own system image, if I am using the standard configuration.

[carstenbauer]

Well, you are still free to reinstall Julia if we add this option. Note that one might be using MKL.jl in combination with a self-compiled julia in which case you would have to recompile entirely.

I'd rather get the default build from julialang.org instead of building my own system image, if I am using the standard configuration.

May I ask why? In what way is the default system image better than a self-built one?

[ViralBShah]

The default system image is built with a great amount of care. So unless I want something out of the ordinary, I personally prefer to use the download provided at julialang.org.

[carstenbauer]

With the new libblastrampoline infrastructure of Julia 1.7 this becomes irrelevant.

[ViralBShah]

I will note that with the LBT stuff, we currently allow loading MKL to override OpenBLAS, but do we need a function to go in the reverse direction? Just restarting Julia will give you OpenBLAS by default (unless you load MKL).

[carstenbauer]

I think that would be great, if only for comparing OpenBLAS and MKL performance when benchmarking a function. For symmetry we might want to add two functions openblas() and mkl()?

[ViralBShah]

It is currently done in the __init__() of MKL.jl. I am trying out BLIS when I come back to this next. I am thinking perhaps a switch_blas API, which in base Julia gets you back to OpenBLAS. MKL.jl, BLIS.jl etc. can then extend it for their own purposes.

@staticfloat - Any thoughts?

[staticfloat]

I'm thinking we should take these lines, wrap them in a function like reset_vendor() and then users can call LinearAlgebra.reset_vendor(). So to switch to MKL, you call using MKL or once it's loaded you can call MKL.use_mkl() or something, and then if you want to reset, you would call LinearAlgebra.reset_vendor().

[carstenbauer]

Can we reopen this to not forget about it? The last few posts support reopening this.

I just stumbled over it again in the context of setting up a centralised Julia depot on a HPC cluster. The depot contains a centralised MKL.jl installation. It also contains a central startup.jl into which I want to put using MKL (as this is what most users want when loading the Julia-Intel module). However, currently, this would force users to use MKL as there is no go_back_to_using_openblas() function. Leaving the using MKL to the user is of course annoying as well (he/she has to potentially type it in every session).

[carstenbauer]

Do we want the function reset_vendor() to live in LinearAlgebra or MKL.jl?

[ViralBShah]

MKL.jl, I think. Will get it in people's hands sooner.

[ViralBShah]

I'm changing my mind on this. I think the reset_vendor() should be in LinearAlgebra so that you can get the Julia default BLAS that shipped. PRs welcome in base!

[ViralBShah]

At this point, given that we have moved to LBT since 1.7, I think the easiest way to reset on 1.6 is to reinstall Julia.

[carstenbauer]

I agree, but for LBT (i.e. Julia >=1.7) we still need a way to revert back to the default BLAS within a session. In fact, I still want a way to switch back and forth. Locally, I have a Julia PR that provides a BLAS.reset_config(). I hope to find the time to finish it soon. Only thing left is a PR to MKL.jl then which adds a MKL.activate() or whatever we want to call it.

[giordano]

I usually do LinearAlgebra.__init__(), or you can call directly what that function does.

[carstenbauer]

Oh I see that we have #90 for tracking these efforts. So let's close this one in favor of #90.

PS: @giordano, yeah, my reset_config() is based on the code in __init__().